Towards Learning Rich Logical Schemas From Natural Language Stories Lane Lawley Gene Kim Lenhart Schubert llawley@cs.rochester.edu gkim21@cs.rochester.edu schubert@cs.rochester.edu cs.rochester.edu/u/llawley cs.rochester.edu/u/gkim21 cs.rochester.edu/u/schubert
Talk Outline 1. A little background 2. Overview of our approach 3. Schema matching 4. Schema generalization 5. Conclusion
Talk Outline 1. A little background 2. Overview of our approach 3. Schema matching 4. Schema generalization 5. Conclusion
Background Schemas are stereotypical patterns of events that can be instantiated into specific stories Schank & Abelson’s scripts , Minsky’s frames ...
Background: uses Inferences from incomplete information: “Skid, crash, hospital” What skidded and crashed? Who was in the hospital? Planning: “I’m hungry. What are some ways I know for someone to get food?”
Background: learning from natural language GENESIS: used database of actions, preconds, effects to abstract causal chains (Mooney, 1991) Cluster verb/subj and verb/obj pairs w/ co-referring args, use SVM for ordering (Chambers & Jurafsky, 2008) LSTMs to find sequences of 5-tuples (verb, subj, obj, preposition, prep arg) (Pichotta & Mooney, 2016)
Talk Outline 1. A little background 2. Overview of our approach 3. Schema matching 4. Schema generalization 5. Conclusion
How does our approach differ? Statistical script learning uses simple event representations… (...and requires a lot of data.) We use episodic logic as our representation language. EL is closer to “surface-form” English than FOL, & much more expressive than 2- or 5-tuples.
How does our approach differ? Symbolic approaches to schema learning have (historically) had bootstrapping issues. - GENESIS (Mooney, 1991) relied on many pre-programmed actions like “ARREST” to understand its stories, which were targeted at adults. These actions were fully specified in FOL w/ preconditions, postconditions, etc. - IPP (Lebowitz, 1983) had a similar pre-programmed action set, & targeted news articles. We target children’s stories and bootstrap with only a set of initial schemas that any 1- or 2-year old child would possess, covering wide sets of actions (do action for pleasure, help someone do an action…). We can also compare two similar stories to find common components of a schema, rather than inferring necessary ones from fully-specified causal chains in one story.
Representation: EL/ULF English: E1 is an episode of Rivka insisting Rivka insisted that she didn’t eat chametz. that there is some E2, and Episodic Logic (EL): E2 is an episode of Rivka not eating chametz, and [E1.sk before Now0], [[|Rivka| insist.v E2 happened at or before E1. (that (some e2 [e2 at-or-before E1.sk] [(not [|Rivka| eat.v (K chametz.n)]) ** e2]))] ** E1.sk)
Representation: EL/ULF English: Underspecified Logical Form (ULF): Rivka insisted that she didn’t eat chametz. (|Rivka| ((past insist.v) (that (she.pro ((past do.aux-s) Episodic Logic (EL): not (eat.v (k chametz.n))))))) [E1.sk before Now0] [[|Rivka| insist.v Tense information not “deindexed” (that (some e2 [e2 at-or-before E1.sk] into explicit episodes like E1 & E2... [(not [|Rivka| eat.v (K chametz.n)]) ** e2]))] ** E1.sk)
( epi-schema ((?x do_to_enable_action.v ?a1 ?a2) ** ?e) ( :Nonfluent-conds !r1 (?a1 action1.n) !r2 (?a2 action1.n) Example !r3 (?x agent6.n)) Schema ( :Goals ?g1 (?x want1.v (that (?x can.md (do2.v ?a2))))) ( :Init-conds ?i1 (not (?x can.md (do2.v ?a2))) ) ( :Steps ?e1 (?x do2.v ?a1)) ( :Post-conds ?p1 (?x (can.md (do2.v ?a2)))) ) ( :Episode-relations !w1 (?e1 same-time ?e) !w2 (?e1 consec ?p1) !w2 (?e1 cause-of ?p1)))
Talk Outline 1. A little background 2. Overview of our approach 3. Schema matching 4. Schema generalization 5. Conclusion
( epi-schema ((?x do_to_enable_action.v ?a1 ?a2) ** ?e) ( :Nonfluent-conds !r1 (?a1 action1.n) !r2 (?a2 action1.n) !r3 (?x agent6.n)) ( :Goals ?g1 (?x want1.v (that (?x can.md (do2.v ?a2))))) ( :Init-conds ?i1 (not (?x can.md (do2.v ?a2))) ) ( :Steps ?e1 (?x do2.v ?a1)) ( :Post-conds ?p1 (?x (can.md (do2.v ?a2)))) ) ( :Episode-relations !w1 (?e1 same-time ?e) !w2 (?e1 consec ?p1) !w2 (?e1 cause-of ?p1)))
; The monkey can climb a tree. ((E1.SK AT-ABOUT.P NOW0) ^ (MONKEY1.SK MONKEY.N) ^ (TREE1.SK TREE.N) ^ ((MONKEY1.SK (CAN.MD (CLIMB.V TREE1.SK))) ** E1.SK)) ; He climbs the tree and gets a cocoanut. ((E2.SK AT-ABOUT.P NOW1) ^ (TREE2.SK TREE.N) ^ ((HE.PRO (CLIMB.V TREE2.SK)) ** E2.SK)) Story from The New McGuffey First Reader
BEST SCHEMA MATCH: ( epi-schema ((?x do_to_enable_action.v ?a1 ?a2) ** ?e) (7 points; 2 from # of consistent temporal constraints, 5 ( :Nonfluent-conds from # of bound variables) !r1 (?a1 action1.n) !r2 (?a2 action1.n) !r3 (?x agent6.n)) ( :Goals (epi-schema ((MONKEY1.SK DO_TO_ENABLE_ACTION.V (KA (CLIMB.V TREE2.SK)) ?g1 (?x want1.v (that (?x can.md (KA (GET.V COCONUT1.SK))) (do2.v ?a2))))) ** ?E) ( :Init-conds (:NONFLUENT-CONDS ?i1 (not (?x can.md (do2.v ?a2)))) !R1 ((KA (CLIMB.V TREE2.SK)) ACTION1.N) ( :Steps !R2 ((KA (GET.V COCONUT1.SK)) ACTION1.N) ?e1 (?x do2.v ?a1)) !R3 ( MONKEY1.SK AGENT6.N) ( :Post-conds ) ?p1 (?x (can.md (do2.v ?a2))))) (:GOALS ( :Episode-relations ?G1 (MONKEY1.SK WANT1.V !w1 (?e1 same-time ?e) (THAT (MONKEY1.SK CAN.MD (DO2.V (KA (GET.V COCONUT1.SK))) ))) !w2 (?e1 consec ?p1) ) !w2 (?e1 cause-of ?p1))) (:INIT-CONDS ?I1 (NOT (MONKEY1.SK (CAN.MD (DO2.V (KA (GET.V COCONUT1.SK)))))) ) BINDINGS: (:STEPS E2.SK (MONKEY1.SK DO2.V (KA (CLIMB.V TREE2.SK)) ) ?X: MONKEY1.SK ) ?A1: (KA (CLIMB.V (:POST-CONDS E3.SK (MONKEY1.SK (CAN.MD (DO2.V (KA (GET.V COCONUT1.SK))))) TREE2.SK)) ) ?E1: E2.SK (:EPISODE-RELATIONS !W1 (E2.SK CONSEC E3.SK) ?A2: (KA (GET.V !W2 (E2.SK CAUSE-OF E3.SK) COCONUT1.SK)) !W3 (E2.SK CAUSE-OF E3.SK) )) ?E2: E3.SK
Talk Outline 1. A little background 2. Overview of our approach 3. Schema matching 4. Schema generalization 5. Conclusion
Generalizing protoschema matches Starting protoschema: “ agent does action to enable action ” New schema: “ monkey climbs tree to get coconut ” We could generalize a bit…. “mammal climbs tree to get fruit”
Generalizing protoschema matches Starting protoschema: “ agent does action to enable action ” New schema: “ monkey climbs tree to get coconut ” We could generalize a bit…. But…. “mammal climbs tree to get fruit” “vertebrate climbs plant to get ingredient”
Two issues (of many...) 1. Combinatorial explosions of possible generalizations 2. Not all schemas are “refinements” of protoschemas
Lazy composite generalization Lazy: We could hold off on generalizing until we see a second, similar story, to narrow the combinatorial generalization space Composite: Incorporate some “incidental” information into a new schema, even when it doesn’t match an existing schema/protoschema
Lazy composite generalization example Simeon and Pedro like to fish. There are fish in their pond. Sometimes they sit on the bridge. They are very nice fish. We will Sometimes they sit on the bank of come and catch them. We will the river. They have poles, long take the long rod, and the hook lines, and little iron hooks. This and line. We must have a bag, morning Simeon caught a large too. It must be strong, to keep fish. Pedro caught many small the fish safe. ones. The boys caught some crabs, too. They use a net to catch the crabs.
Roles Simeon and Pedro like to fish. There are fish in their pond. Sometimes they sit on the bridge. They are very nice fish. We will Sometimes they sit on the bank of come and catch them. We will the river. They have poles, long take the long rod, and the hook lines, and little iron hooks. This and line. We must have a bag, morning Simeon caught a large too. It must be strong, to keep fish. Pedro caught many small the fish safe. ones. The boys caught some crabs, too. They use a net to catch the crabs.
Shared roles Simeon and Pedro like to fish. There are fish in their pond. Sometimes they sit on the bridge. They are very nice fish. We will Sometimes they sit on the bank of come and catch them. We will the river. They have poles, long take the long rod, and the hook lines, and little iron hooks. This and line. We must have a bag, morning Simeon caught a large too. It must be strong, to keep fish. Pedro caught many small the fish safe. ones. The boys caught some crabs, too. They use a net to catch the crabs.
What about fish/crab, pond/river? Simeon and Pedro like to fish. There are fish in their pond. Sometimes they sit on the bridge. They are very nice fish. We will Sometimes they sit on the bank of come and catch them. We will the river. They have poles, long take the long rod, and the hook lines, and little iron hooks. This and line. We must have a bag, morning Simeon caught a large too. It must be strong, to keep fish. Pedro caught many small the fish safe. ones. The boys caught some crabs, too. They use a net to catch the crabs.
What about fish/crab, pond/river? LCH(FISH, CRAB) = ANIMAL LCH(POND, RIVER) = BODY_OF_WATER LCH(SIMEON, PEDRO, WE, BOYS, THEY) = PERSON
Recommend
More recommend
